AIBO Common World Model

1. AIBO Common World Model Thomas Jellema Stephan Kempkes Emanuele Venneri Rob Verkuylen

2. Introduction • Started in 1980’s • Not too many practical applications • Testbeds • Foraging and Coverage • Multi-target Observation • Box pushing and object transportation • Exploration and flocking

3. Introduction • Multi-Robot • Global task which cannot be achieved by a single robot • Higher performance by using multiple robots • Multi-Robot vs. Multi Agents • Real environment (uncertainty)

4. Taxonomy Cooperation

5. Taxonomy Cooperation Knowledge • No knowledge of other robots, no communication • Communication possible through stigmergy • Mainly used for foraging or box pushing

6. Taxonomy Cooperation Knowledge Coordination • No communication protocol • Taking into account other robot’s actions • “Follow the leader” object transportation

7. Taxonomy Cooperation Knowledge Coordination • No communication protocol • Reduce interference • Aircraft free-flight mode (small detour allowed)

8. Taxonomy • Using communication protocol • Leader chosen in advance • High computational demand for leader Coordination Organization

9. Taxonomy • Using communication protocol • Leader chosen dynamically • More robust than strongly centralized organization Coordination Organization

10. Taxonomy • Using communication protocol, autonomous actions • Robust to communication failure and robot malfunctioning • Broadcast of Local Eligibility Coordination Organization

11. World Model • Robots maintain individual world model • Contains perceptions of state of world • Soccer: position, heading, ball, teammate and opponent positions • Use all individual models for global world model • Shared perception to minimize sensor reading error

12. Previous work • AIBO team with global world model • Difference in timestamps; no Kalman filtering • High latency; only use shared ball position when ball cannot be located • High error in vision; use teammate’s own estimated positions

13. Proposal To create a flexible communications platform for inter AIBO information exchange to the extend to create a Common World Model.

14. Proposal - cont Strategy Common World Model Remote Event … Messages Protocol Physical

15. References • Dave Touretzky and Ethan Andrew of CMU • Add multi-connection support to Wireless • Help with Messaging Design • German Code has form of CMW • Gal Kaminka of Bar-Ilan • Extensive Multi-Robot Research with AIBO’s

16. Approach • Research Previous Work • Create Multi AIBO Message Layer • Add multi-connection support • Register AIBO’s • Subscribe to Events • Expand current WorldModel to facilitate Common World Model • Experiment with Common World Model

17. Message Layer Design • AIBO’s registered with each other. • Static or Dynamic(broadcast) • EventRouter instance for each remote AIBO • AIBO’s subscribe to receive event types. • If no AIBO subscribes the originating AIBO won’t send the events or receive all events and just act on the ‘interesting’ ones.

18. Advantages • Subscribing to events on other robots would be as simple as subscribing to events on the local machine. • Flexible message layer usable for various multi-robot applications. • Common World Model is just one application of the Message Layer. • Updating the CWM in an AIBO would be the behavior to the event containing the WorldModel of another AIBO.

19. Planning

20. Questions